Headline in Arial Bold 30pt - HWSW
21
Headline in Arial Bold 30pt Taking the New Road Taking the New Road John Masters John Masters Marketing Director, EMEA Marketing Director, EMEA
Transcript of Headline in Arial Bold 30pt - HWSW
Headline in Arial Bold 30pt
Taking the New RoadTaking the New RoadJohn MastersJohn Masters
Marketing Director, EMEAMarketing Director, EMEA
HPC Issues : 2006 feedback
• From Reliance on Clock-Speed to Multicore
• Large-Cluster Issues : Power, Reliability, Weight
• Global shared memory but at Lower Cost
• Vendor viability
HPC Issues : 2006 feedback
• From Reliance on Clock-Speed to Multicore
• Large-Cluster Issues : Power, Reliability, Weight
• Global shared memory but at Lower Cost
• Vendor viability
Change in Processor Progress : Clock-Speed to Multicore
LogScale
01980
Processor ChipAggregate Performance2x every 1.5 to 2 yearsPentium4, 3200MHz
Intel® 8086, 5MHz
Asymptotic
1990 2000 2010
HPC Issues : 2006 feedback
• From Reliance on Clock-Speed to Multicore
• Large-Cluster Issues : Power, Reliability, Weight
• Global shared memory but at Lower Cost
• Vendor viability
Packaging Strategy : Energy Efficiency Carlsbad ‘07 : 1.05 PF Peak in 192 Racks
• ~100K cores (2.66 GHz)• ~6K 24-port IFBDDR switches• ~55’ x 60’ footprint• 6 MW (80% derate)
• 5.8 MW, 1PF peak, ’07• 4.5 MW, 1PF peak, ’08
8 R
anks
24 Racks / Rank
0
1
2
3
4
5
6
2007 2008 2009 2010 2011
MW per PF peak
Packaging Strategy : Density & Reliability Carlsbad ‘07 : Cable-Free IRU (much reduced cabling)
12 Groups = 36 Racks (3 Racks / Group)
8 R
anks
• 1.5 PF peak• Normally = 55,296 cables
• Via cable-free IRU= 18,432 cables
IRU(128 cores; no cables)
Carlsbad ‘07 : Density => Weight
L1 Display
L1 Display
L1 Display
L1 Display
L1 Display
L1 Display
L1 Display
L1 Display
(16) AtokaP Blades
• 2,050 lb• 24” Rack• 246 lb / ft2
• 2,050 lb• 19” Rack• 275 lb / ft2
(250 lb / ft2 limit)L1 Display
L1 Display
L1 Display
L1 Display
L1 Display
L1 Display
L1 Display
L1 Display
(16) AtokaP Blades
Copper : Weight becoming significant
Picture credit : LRZ
Copper : Weight becoming significant. Why?
LogScale
01980
Processor ChipAggregate Performance2x every 1.5 to 2 years
Copper Wire-pairBandwidth
2x every 3 to 4 yearsAsymptotic
1990 2000 2010
HPC Issues : 2006 feedback
• From Reliance on Clock-Speed to Multicore
• Large-Cluster Issues : Power, Reliability, Weight
• Global shared memory but at Lower Cost
• Vendor viability
Interconnect Strategy : Reduce Cost & Increase Capability
• NUMAlink4 (Today)
– Custom copper cable
– Custom signalling
– Custom protocol
• NUMAlink5: hw extension of IFB (’09)
– COTS Infiniband12x copper cable
– COTS serdes
– Custom protocol (higher capability)
$450 (5m) $150 (5m)
SGI Server Roadmap
CY 2006 2007 2008 2009
Altix 4700SHub2, NL4
Montvale
Big NodesccNUMA
to 512S nodesTB!PB GAM
UltravioletUVH, NL5
Tukwila / Beckton
Altix 4700SHub2, NL4Montecito
1.5TB/Rack (0.8TF,51GF/W)
CarlsbadEDR Optimized
Clovertown5.5TF/Rack 136GF/W
CarlsbadEDR Optimized
Harpertown7.2TF/Rack 179GF/W
Cluster2S nodes
10sGB GAMAltix XE210/240
Woodcrest
Altix XESanta Fe/Dixon
Clovertown
--
Tigerton ! Dunnington ! Bloomfield
Interconnect Strategy : Reduce Cost & Increase Capability Shub2 (Today) to UV-Hub (2009)
NUMAlink to Other Nodes
[S]Intel
Today’s Socket
PI
MI
NI
Globally Addressable Memory
SGISHub2180 nm
PI
MI
NI90 nm
[S]Intel
CSI Socket
[V] GRU
[P] AMU
HPC Issues : 2006 feedback
• From Reliance on Clock-Speed to Multicore
• Large-Cluster Issues : Power, Reliability, Weight, IO-Data
• Global shared memory but at Lower Cost
• Vendor viability
The “New SGI” Financially Healthy
• Financial / Re-Engineering Complete– Emerged from Chapter 11 one month earlier than original plan – Recapitalised the Company– Eliminated $250M legacy debt– Bolstered our cash position
• New Leadership Team, New Board of Directors
• Costs Reduced $150M Annually
• Stabilised the Business
• Most Complete Product Portfolio in Company’s History
• Focus on Growth Initiatives
SGI® Altix® Dual-Core ServersAltix 450 and Altix 4700
SGI® Altix® 450 SGI® Altix® 4700
Revolutionary Architecture Leads Next-Wave of Blade-based Platforms
Innovative Blade-to-NUMALink4 Design: Provides unprecedented versatility, density, power efficiency, performance
Investment Protection: Processor-only upgrade to future Intel® Itanium® 2 processors or expansion via new blades
Enables Multi-Paradigm Computing: Enhanced integrated computational resources that can be seamlessly accessed as application requires
With the Scalable, Reliable Performance on Linux that Altix is Known for
“With the SGI Altix we’re significantly increasing our capacity for large-scale computations requiring large memory, problems that are even beyond the scope of our existing system, which is hardly small. We’ve run some benchmarks, and jobs that were taking about 50 to 60 hours on the HP supercomputing system actually took about 2 to 6 hours on a similarly configured SGI Altix. I consider the Altix to be a keystone in our infrastructure.”- Dr. Gordon Springer, Scientific Director, Bioinformatics Consortium, University of Missouri
SGI® Altix® XE1200 Clusters
• Competitive price/performance• Dual Intel® Xeon® Processors• Linux® OS-based• Full factory integration for plug ‘n’ play implementation in target markets • Technologically advanced architecture (energy efficiency, enhanced software environment)• Solid roadmap for future differentiation based on extended performance and scalability enhancements
“IDC research studies show that end users frequently find large clusters very complex to manage and they require significant effort to setup and optimize. SGI's approach is to integrate application-optimized clusters in the factory and to support these systems with the same team that has serviced its supercomputers and storage products.”
- Earl Joseph, vice president for IDC's Technical Computing Systems practice
SGI® Altix® XE1300 Clusters
• Dual socket with Dual/Quad core Xeon based processors
• 2 servers in 1U ( ‘server’ = mobo, disk, 2 x gigE, IB …)" Power per server reduced " Half the number of Power Supply Units
• Infiniband integrated onto motherboards Dramatically reduced cost for high performance interconnect
• 512 cores + 23kW in 1 rack
ROADMAP : XE1300 + Tighter integration = ‘Carlsbad’
HPC Issues : 2006 feedback
• From Reliance on Clock-Speed to Multicore
• Large-Cluster Issues : Power, Reliability, Weight
• Global shared memory but at Lower Cost
• Vendor viability
Headline in Arial Bold 30pt
Questions?
